Fides: agentic paper reproduction

An AI lab that reproduces published papers from their own equations and data. It freezes verifiable numeric targets before any code exists, re-implements the method, runs physics checks (gates), and issues a verdict it cannot write by hand. An adversarial critic then reviews every run. Runs on our own fleet models (GLM-5.3-Flash, DeepSeek-V4.1-Flash).

2026

How a reproduction works

  1. Plan the reproduction from the paper's own equations and data.
  2. Freeze and seal the targets: verifiable numbers, fixed before any code exists. The lock is hash-checked, and edits to a sealed lock are blocked.
  3. Implement the method one committed step at a time.
  4. Gates before targets: physics checks must pass before any target is compared.
  5. Verdict from verify.py, never written by hand: REPRODUCED, REPRODUCED_LOOSE, DIVERGENT, NOT_REPRODUCIBLE or PENDING, each with an evidence label.
  6. Adversarial audit: an auditor reads the git history for tuning to the target.

Status

Private beta, by invitation.

Example reproductions

None published yet. Each example appears here once it has been reviewed.

Team

Dirk Englund · PI
← Back to Dirk Englund